In these years attention is being given to power saving techniques aiming at reducing heat generation in information apparatuses. The following documents are considered:                [Non-Patent Document 1) “Intel® Pentium® Processor Enhanced Intel SpeedStep Technology”, Webpage URL://support.intel.com/support/processors/mobile/pm/sb/CS007981.htm        [Non-Patent Document 2] M. Elnozahy, M. Kistler, and R. Rajamony, “EnergyConservation Policies for Web Servers”, USITS 2003        [Non-Patent Document 3] N. Kandasamy, S. A. Abdelwahed and J. P. Hayes, “Self-Optimization in Computer Systems via Online Control: Application to Power Management”, In Proceedings of the International Conference on Autonomic Computing, 2004        [Non-Patent Document 4] “WebSphere® Extended Deployment Edition”, Whitepaper,WebpageUIRL: 6.ibm.com/jp/software/websphere/ft/was/xd/pdf/whitepaper.pdf        
A technique for changing the operating frequency or voltage of central processing units has been used (for example see Non-Patent Document 1). In order to reduce energy consumption in a Web server by using this technique, another technique has been proposed that controls the operating frequency of the Web server on the basis of a 90% response time of the Web server (see Non-Patent Document 2).
The term “90% response time” as used in that document refers to the time required for a Web server to respond to 90% of the processing requests sent to the Web server. The 90% response time is used as a measure of how quickly the Web server can perform processing. According to the technique, the 90% response time is measured at regular time intervals. If the measured value is less than a target value, the operating frequency is reduced; if the measured value exceeds the target value, the operating frequency is increased. This allows the target to be achieved while preventing excessive power consumption.
A technique that controls the operating frequency of a Web server in accordance with the number of processing requests to the Web server has been proposed (see Non-Patent Document 3). According to the technique, the trend in recent variations in the incoming rate of processing requests is determined from the difference between the incoming rate of processing requests previously measured and the incoming rate of processing requests currently measured and a predicted value of the feature average response time is calculated on the basis of this trend. Then, the lowest operating frequency that can maintain the average response time less than a target value is set. Thus, if variations in the processing request incoming rate show a given trend, appropriate control can be performed.
The present invention solves the following problems. Because the 90% response time is determined on the basis of response times over a period between a predetermined past time and the present time, there is a time lag between the change of the operating frequency and a change in the 90% response time. Therefore, the 90% response time in the technique described in Non-Patent Document 1 can fluctuate around the target value. The technique described in Non-Patent Document 1 enables operating frequency control at very short time intervals by including the function of monitoring TCP/IP packets to measure the 90% response time and the function of controlling the operating frequency in the kernel of an operating system. Accordingly, the range of fluctuation is relatively narrow.
However, because it is difficult to identify the types of processing requests at the level of TCP/IP packets, only one response time goal can be addressed for all application programs running on a Web server. This is unrealistic and it is desirable that a response time goal be set for each application program. In fact, the technique described in Non-Patent Document 4 allows a response time goal to be set for each individual application program.
To address more than one target values, processing requests must be monitored at a higher level such as the HTTP. HTTP requests are monitored in user space. Measurement of the response time and operating frequency control in user space use a greater overhead than in kernel space. Accordingly, if operating frequency control is performed at short time intervals, the performance of Web applications would be decreased significantly.
If a response time goal is provided to a workload manager external to a Web server as in the technique described in Non-Patent Document 4, the operating frequency of the Web server must be controlled externally through a network. Accordingly, further overhead is consumed. Therefore, it is difficult to set time intervals for controlling the operating frequency to a small value on the order of several tens of seconds and the operating frequency must be controlled at longer time intervals.
If the technique described in Non-Patent Document 2 is applied to such a system without modification, time intervals for measuring the 90% response time will become considerably long. As a result, the 90% response time can fluctuate in limits far away from the target value. This holds true also for the technique described in Non-Patent Document 3. If operating frequency control is performed at long time intervals, the average response time can fluctuate.